Anthropic researcher quits over self-improving AI risk
A pretraining researcher announced his exit from Anthropic on September 9, warning that labs are racing to self-improving superintelligence unprepared.
Symbolic image: a person unplugs a network cable from a running GPU server rack at night while the status lights keep blinking.
Anthropic pretraining researcher Jacob Coxon resigned on September 9, 2026, saying the industry is racing toward self-improving superintelligence without any plan to control it.
At a glance
- Jacob Coxon made his Anthropic resignation public in a social media post on the evening of September 9, 2026.
- He had spent three years on pretraining research, first at OpenAI and later at Anthropic.
- His charge: labs are racing straight to self-improving superintelligence and gambling with everyone's lives.
- Colleague Evan Hubinger puts the chance that AI kills all humans within the next decade at more than 10 percent.
- Anthropic had not responded to a request for comment, according to TechCrunch.
Anthropic pretraining researcher Jacob Coxon resigned on September 9, 2026 and posted his reasoning publicly the same evening: he believes AI labs are heading straight for self-improving superintelligence with no plan for keeping it under control. Coxon had spent three years on pretraining research, first at OpenAI and later at Anthropic.
An insider's objection
His central complaint is about pace rather than capability. As he describes it, the danger is well understood inside the labs; the problem is that everyone involved is locked into a race to get there first. He accused the industry of “gambling with our lives” — a line aimed squarely at colleagues, not at outside critics.
A colleague puts a number on it
Anthropic researcher Evan Hubinger backed the substance of the warning and went further, putting the odds that AI kills every human within the next decade at more than 10 percent. He also conceded there is no finished plan for solving alignment for superintelligence. Numbers like that are individual estimates, not measurements, and are worth reading that way.
What is confirmed, and what is not
The resignation, the stated reasoning and Hubinger's remarks all come from TechCrunch's September 9, 2026 report. Anthropic had not commented by the time that story was published. Coverage from other newsrooms was not reachable for us, so nothing from those reports has been checked or used here.
Why this is bigger than one exit
Anthropic built its public identity on safety research, which makes a safety-motivated departure more awkward for it than it would be for a rival lab. The question on the table is no longer whether the models are useful, but whether the schedule is defensible. Whether this stays a single news cycle depends on how many colleagues follow him out the door.
FAQ
Why did the Anthropic researcher quit?
Jacob Coxon says AI labs are racing to build self-improving superintelligence without a workable plan to control it, and he did not want to keep contributing to that race.
What is self-improving AI?
AI systems that drive their own further development. Coxon argues the speed of that race is the danger, because no reliable control method exists yet.
What odds do Anthropic researchers give of AI killing everyone?
Evan Hubinger puts it at more than 10 percent within the next decade. That is his personal estimate, not an official company figure.
